Vertical AI

For an AI startup selling to businesses, a paid pilot should have defined criteria for moving into regular use Before it begins, the team and customer should agree on: ◻️ The task being tested and the expected result ◻️ The time and cost of the existing process ◻️ Acceptable error rates and required human review ◻️ Who controls the budget and makes the renewal decision During the pilot, the startup should record the full cost of delivery: implementation hours, model usage, retries, support and manual corrections Onboarding and custom development fees should be reported separately from recurring product revenue This helps establish how much income depends on additional work for each customer At renewal, review whether the customer uses the product regularly, achieves the agreed results and accepts pricing that covers delivery costs Any discount or additional support needed to secure the renewal should remain visible in the assessment For an incubator, this evidence should inform the next funding decision If each deployment requires extensive custom work, the budget should account for that delivery model If deployments become more consistent, the team can test whether acquiring additional customers produces sustainable margins The next review should compare those expectations with actual renewals, implementation hours and margins across subsequent customers

◻️ On 19 October, NorForge documented more than 1,100 attack projects involving AI agents across 14 exposed dashboards linked to one activity cluster The figure counts projects, not confirmed breaches ◻️ In January and February, Amazon documented an AI-assisted campaign that compromised over 600 FortiGate devices across more than 55 countries An attacker with limited technical skills gained entry through weak credentials and exposed management interfaces AI helps attackers research targets, build tools and adapt their methods. Smaller teams can attempt more attacks with less manual work Agents deployed inside companies can also be targeted Malicious instructions hidden in a document or webpage can cause an agent to expose data or misuse its access to other systems As companies give agents access to customer records, code and payments, they need to fund the security work that comes with it Permissions must be enforced outside the model Sensitive actions need separate authorisation. Teams need logs, monitoring and a way to revoke access quickly Security testing must continue as software changes. AI can help defenders find vulnerabilities and check whether fixes work For founders and incubators, due diligence should establish who owns security, which systems agents can access, how incidents are handled and whether the operating budget covers that work

AI Is Lowering the Cost of Cyberattacks Artificial intelligence helps attackers research targets and develop tools with less technical knowledge and manual work In its 2026 Global Threat Report, CrowdStrike reported an 89% year-on-year increase in the activity it observed from attackers using AI Two documented cases 👇

Public token fundraising has recovered unevenly. In CryptoRank’s dataset, the average disclosed public token sale raised approximately: ◻️ 2021: $1.50 million ◻️ 2023: $460,000 ◻️ 2025: $2.08 million ◻️ 2026 YTD, through 10 June: $3.15 million. Activity remained volatile in 2026 CryptoRank reported just $8.4 million in total token-sale fundraising in August. The first week of October recorded the highest weekly total since February Capital remains concentrated Coinbase’s analysis of 962 projects with documented token-sale proceeds from 2021 to September 2026 found that 130 projects accounted for approximately 75% of proceeds This broader sample includes public and private sales beyond launchpads The sector mix also changed. Within that sample: ◻️ Infrastructure, L1/L2 networks and DeFi represented 64.5% of sales in 2025 ◻️ AI’s share rose from 1.3% in 2021 to 13.2% in 2025 ◻️ Gaming and metaverse sales fell from 75 in 2021 to 17 in 2025 The January–September 2026 subset contains only 20 documented sales, with recent disclosures incomplete. It does not support a reliable update to those sector shares Selected CoinList offerings in early 2026 illustrate three fundraising themes: ◻️ AI infrastructure finance: USD.AI - lending against GPU collateral ◻️ Confidential infrastructure: Zama - computation on encrypted data ◻️ Integrated DeFi: Flying Tulip - trading and lending within one protocol

Launchpads need to prove their value before and after a token sale Only 13 public token sales closed in May 2026, the lowest monthly figure since late 2020 Active institutional venture investors fell from 767 in Q3 2025 to 333 in Q3 2026 Returns have given buyers little reason to return: by December, 84.7% of 118 tracked launches from 2025 traded below their initial valuation We see three directions for launch platforms: ◻️ Selection and support before the sale. Teams undergo formal review and receive help with tokenomics, market entry and fundraising preparation. Some programmes also provide initial funding ◻️ Regulated sales. In the EU, professionally placing tokens on an issuer’s behalf is a regulated service under MiCAR, requiring the relevant authorisation ◻️ Control over capital after the sale. Some platforms hold proceeds in treasuries that teams cannot move independently Monthly budgets cover operations, while additional spending requires governance approval A launchpad seeking to rebuild trust should explain why it selected a team, what support that team receives and who controls the money after the raise